AI trained on 3500 years of games finally beats humans at Dota 2 | Tech News
They say 10,000 hours makes an expert, but for video-game playing AIs much more is needed. After playing thousands of years’ worth of the video game Dota 2, artificial intelligence is now able to beat the world’s top amateurs.
Dota 2 is a multiplayer battle arena game. Two teams of five take on each other across a large map, defending their own base whilst trying to attack their opponents. It is the game that has the largest prize money of any e-sport, so there are many professional players able to play at a very high level.
To teach AI to play the game, research lab OpenAI used a technique called reinforcement learning. The AI is thrown into the game and learns to play through trial and error. It plays games against itself, and the programmers set different reward criteria that it tries to optimise.
But this is no short-cut to perfection. To reach its current level, the AI had to play 180-years’ worth of games every day for 19 days. It was then able to beat the top 1 per cent of amateurs at the game.
It is still a “big jump” for AI to become the best Dota 2 player in the world, says Greg Brockman at OpenAI. “If a friend came to you and said I’ve beat some semi-pros at something, you wouldn’t expect them to go on and win the world championship,” he says.
However, later this year and after some more training OpenAI will pit their AI against the world’s best Dota 2 players at The International – the biggest e-sports event.
Previously AI has mastered games like chess and go. These are both complex games with many different possibilities, but they are lots of aspects of the real-world that are still missing says Brockman.
For example, for an AI to be successful in the real world it needs to interact in real-time, not by taking turns like in most board games. And similarly, it may not be able to see the whole picture all of the time so needs to be able to make decisions based on imperfect information. Both of which are better reflected in Dota 2 than in chess or go.
In principle, this should mean that AI that can master Dota 2, should be better at mastering tasks in the real-world. “We’ve shown that AI can learn very intelligent behaviours and we want to find other domains to apply this in,” says Brockman.